A bachelor thesis project about autonomous car maneuver around roundabout using reinforcement learning with deep q network.
Conference paper published at: https://ieeexplore.ieee.org/document/10221077
Two kinds of roundabouts are used in this research.
- Roundabout with Intersection
- Roundabout without Intersection
The camera sensor was obtained, and then we did the semantic segmentation using CARLA's built-in function.
->
RGB -> Semantic Segmentation
Two kinds of camera setup are used in this research.
- Segmented and Grayscaled
- Segmented and retouched
There are three actions that the agent can do.
- Forward Throttle = 1; Steer = 0;
- Forward Left Throttle = 1; Steer = -1;
- Forward Right Throttle = 1; Steer = 1;
The reward functions are different for each kinds of roundabout.
- Angle Deviation.
The value is
Reward1 = 1/alpha
. Alpha is the difference in angle from the target line and the agent angle.
- Distance Deviation.
The value is
Reward2 = 1/distance_deviation*10
. Distance deviation is the distance from agent to the target line in m.
-
Collision. The value is
Reward3 = -1
. Collision event triggered when the agent touched another object. -
Agent too far. The value is
Reward2 = -0.5
. Agent too far event triggered when the distance between the center of roundabout and the agent is more than 30 meters. -
Total Reward:. The total reward is
Reward = Reward1 + Reward+2 + Reward3 + Reward4
- Angle Deviation.
The value is
Reward1 = 1/alpha
. Alpha is the angle difference between the agent's direction and the direction from the agent to the nearest waypoint + 5 from the agent.
-
Collision The value is
Reward2 = -1
. Collision event triggered when the agent touched another object. -
Total Reward: The total reward is
Reward = Reward1 + Reward+2
Using reinforcement learning with deep q network.